A Novel Deep Temporal Feature Enhanced Just‐in‐Time Learning Framework for Predicting Rare Earth Component Content

ABSTRACT Real‐time online detection of rare earth element component contents is a crucial link in ensuring the stable production of the rare earth extraction and separation industry and improving the quality of rare earth products. The traditional methods for predicting the content of rare earth element components based on just‐in‐time learning fail to fully explore the temporal characteristics of process variables in the rare earth extraction. To address this issue, this paper proposes a just‐in‐time learning framework enhanced by deep temporal features to achieve real‐time prediction of rare earth element component content. First, a deep long short‐term memory network autoencoder (DeepLSTMAE) is constructed to deeply extract the temporal dynamic features of process variables in rare earth extraction through offline pre‐training. Second, the extracted implicit temporal dynamic features are integrated with the traditional similarity measurement methods in just‐in‐time learning to establish a spatio‐temporal similarity criterion. Based on this criterion, a local sample set for just‐in‐time learning is selected to establish an online local model. Then, the least squares support vector machine (LSSVM) is used to construct an online local model for the selected similar samples, enabling online dynamic prediction of component contents. Finally, through simulation verification using actual data from rare earth production sites and comparative analysis with other online prediction methods for rare earth component contents based on just‐in‐time learning, the results demonstrate the effectiveness and superiority of the proposed method.